diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
new file mode 100644
index 00000000..022d40cc
--- /dev/null
+++ b/docs/ROADMAP.md
@@ -0,0 +1,111 @@
+# Roadmap and design notes
+
+Forward-looking design decisions that are not yet work in flight. Each entry is
+a recommendation with enough reasoning and numbers to act on — or to decide
+"not now" without re-deriving the analysis.
+
+## Engine-side headless VT per session (D6, WI-130)
+
+*Spike, 2026-09-09. Part of the terminal-attach-budget initiative (WI-121).*
+
+### Question
+
+The engine keeps only **raw PTY bytes** in a per-session ring
+(`scrollback.rs`), and every client rebuilds a terminal by re-parsing those
+bytes into xterm.js on attach. Should the engine instead run a **headless
+terminal emulator per session**, so a cold attach ships *screen state*
+(tmux-style) — a serialized screen plus N lines of scrollback — instead of byte
+history?
+
+### What the budget work already fixed
+
+Before committing to a VT, weigh it against what shipped in this initiative:
+
+- **F1 (WI-125)** bounds *every* replay to the tail budget — the 4 MiB aged-out
+  flood is gone; a cold or stale attach now ships at most the budget
+  (default 1 MiB), ground-state aligned.
+- **F3 (WI-127)** stops retained tabs from time-slicing the active pane's
+  parser on reload.
+- **F5 (WI-129)** persists a serialized xterm screen client-side, so a *reload*
+  restores in one write with no raw re-parse, and (once F4 lands) a *switch*
+  keeps the socket open and never re-streams.
+
+So the two operator symptoms (switch re-stream, slow fresh open) are addressed
+by bounding and by client-side serialization, **without** an engine VT. The VT
+is only worth its cost if a residual problem remains after these ship and are
+measured on prod — see "When to revisit".
+
+### Crate options
+
+| Crate | What it is | Fit for a screen-snapshot server VT |
+|---|---|---|
+| `vt100` | A minimal pure-Rust parser that maintains a screen grid + scrollback, no rendering. | **Best fit.** Smallest surface and memory; exposes the cell grid directly, which is all a snapshot needs. Serializing its screen to an xterm-compatible escape stream is a bounded amount of new code. |
+| `alacritty_terminal` | The terminal model behind Alacritty. | Heavier per-cell model and an API shaped around a GPU renderer's needs; more memory and more moving parts than a snapshot server wants, and version churn tied to Alacritty. |
+| `termwiz` | WezTerm's terminal library. | Full-featured (its own line/cell/attribute model, image protocols); the largest surface of the three. Overkill for "hold a screen and emit a snapshot". |
+
+### Cost estimate (per session, at 8 busy sessions)
+
+No live measurement was taken (an agent session cannot reach a prod stack; use
+H1's load session — `load_session_command` in the engine integration tests — to
+measure before adopting). Grounded estimate for `vt100`:
+
+- **Memory.** A cell is a codepoint + attributes ≈ 8–16 B. A 200-col grid with
+  50 visible rows + 1000 scrollback lines ≈ 210 000 cells ≈ **1.7–3.4 MB per
+  session**, i.e. **~14–27 MB across 8 sessions** — *on top of* the existing
+  4 MiB raw ring per session unless the ring is then shrunk. The ring cannot be
+  dropped entirely: warm `resume_from` deltas and the history archive still read
+  raw bytes.
+- **CPU.** The VT must parse every byte the PTY produces, on the hot reader
+  path. The client already parses at ~5–6 MB/s into xterm; a Rust grid VT is
+  faster (tens of MB/s) but it is now paid **once per session, always**, not
+  once per attach. For a chatty agent session this is continuous cost the
+  raw-ring design does not have.
+
+### Wire shape (if adopted)
+
+Add a snapshot kind rather than changing the existing one:
+
+```json
+{"type":"snapshot-start","kind":"screen","session_id":"…","scrollback_pos":N,"reset":true}
+```
+
+followed by a server-produced, xterm-compatible escape stream that redraws the
+screen and the last N scrollback lines, then live bytes as today. `resume_from`
+still selects the byte-delta path (unchanged); `kind:"screen"` is the cold /
+aged-out path only. The history archive is untouched (it keeps raw bytes).
+
+### Fidelity
+
+A server VT and the client xterm are two independent emulators; they can
+disagree (edge cases in wide chars, unusual SGR, DEC private modes). The
+snapshot is only correct if the server's redraw reproduces on xterm what the
+program intended. The H2 corpus + fidelity harness is the right place to prove
+this: parse each corpus in the candidate crate, emit the screen escape stream,
+write it into headless xterm, and compare `serialize()` against xterm fed the
+raw bytes. Adopt only if that comparison is clean across the corpus.
+
+### Mobile
+
+The prize on mobile would be dropping the IndexedDB cache entirely: if a cold
+attach always ships a small screen snapshot, the phone need not persist raw
+scrollback. But F5 already gives mobile a one-write restore from a *bounded*
+serialized cache, so the marginal win is "no client cache at all" versus "a
+small, capped client cache" — modest.
+
+### Recommendation: **not now**
+
+Ship and measure F1–F5 (and F4) on prod first. The engine VT doubles per-session
+memory and adds continuous CPU on the PTY hot path to solve a problem the
+bounding + client serialization work already targets. Revisit **only** if, with
+those shipped and measured:
+
+1. a first cold attach on mobile is still too slow *and* the ≤1 MiB tail is the
+   bottleneck (not network), or
+2. keeping any client-side scrollback cache proves untenable (storage, privacy),
+   or
+3. a new requirement wants server-authoritative screen state (e.g. server-side
+   search over live screens, or thumbnails).
+
+If revisited, the first step is a `vt100`-based measurement against H1's load
+session at 8 sessions, then the H2 fidelity comparison above — no new work items
+are filed now.